Sie sind auf Seite 1von 38

14-1

Seismic Inversion and AVO


applied to Lithologic Prediction

Part 14 - Combining geostatistics and
multiattribute transforms
A channel sand case study
14-2
Introduction
In this talk, we will look at a new approach
to integrating well log and seismic data,
which involves post-stack inversion,
geostatistics, and multiattribute transforms.
This method will be applied to data slices
extracted from multiple 3D volumes.
I will illustrate this approach using the
Blackfoot dataset from Alberta, Canada.
14-3
The Blackfoot survey
This map shows the location of the Blackfoot survey area, with the portion
used in this study outlined in red. The objective, a Glauconitic channel
within the Lower Cretaceous Mannville formation, is shown running north-
south on the map. The survey was recorded in October, 1995.
Alberta
Calgary
N
Channel
14-4
Base map from the Blackfoot survey
This map shows the 12 wells in the seismic survey area, and Xline 18.
Note that we have rotated the map from the previous display.
Xline 18
N
14-5
Correlating the logs with the seismic data
This figure shows the correlation of well 14-09 with the seismic data, where
the synthetic trace is in blue the and the seismic trace is in red. The sonic
and porosity logs are on the right. The top and base of sand are also shown.
14-6
Seismic line from volume
This figure shows Xline 18 from the seismic volume, showing correlated
sonic logs from two intersecting wells, and the picked channel top.
14-7
Line from inverted seismic volume
This figure shows Xline 18 from the inverted volume. The color key
indicates impedance.
14-8
Acoustic impedance slice
This map shows the arithmetic average of the acoustic impedance
over a 10 ms window below the channel top event. Notice the
channel in green (low impedance) on the left of the map.
Channel
14-9
Wells showing average porosity
This map shows the average porosity over the zone of interest at
each well. Notice the high porosity (purple) values in the center left.
High
Porosity
14-10
Initial crossplot
This is the crossplot between the well porosities and acoustic impedance
values. The red line is the regression fit, and the correlation is 0.65.
14-11
Regression applied to inversion slice
This map shows the application of a the regression line from the
previous slide to the inversion slice. Unfortunately, the wells do not tie!
Channel
14-12
Map-based geostatistics
Map-based geostatistics involves producing three types
of maps:
Optimal maps (Best Linear Unbiased Estimates)
from sparse well data (kriging).
Maps that incorporate both sparse well data and a
secondary seismic attribute (cokriging and kriging
with external drift, or KED).
Conditional simulations of a range of possible
scenarios.
Today, we will focus on the kriged and cokriged maps.
Both of these maps are based on the variogram.
14-13
Kriging and the spatial covariance
Kriging involves creating map values (f
0
) that are a
weighted sum of the input well porosities (f
1
to f
N
):
N N 2 2 1 1 0
w w w f f f f
The weights are determined by solving a set of linear
equations involving the spatial covariance, C(h
ij
), where h
ij

is the distance between the i
th
and j
th
porosity values.
The covariance is derived from the variogram, which is
calculated by fitting a smooth function to values derived
from the well values.
The variogram from the wells and impedance slice are
shown in the next slide.
14-14
Variograms
(a) The figure above shows the
variogram from the 12 wells on
the map.
(b) This figure shows the seismic-to-
seismic variogram. Using the
Markov-Bayes linear assumption, we
can use this variogram for all of our
maps.
14-15
Kriged result
This map shows the result of applying kriging to the 12 wells within the
3D seismic survey, using the scaled seismic-to-seismic variogram.
Channel
14-16
Estimating the error
To see the error associated with kriging, we usually
display the error variance. But this is simply the
theoretical error, and will go to zero as the variance of
the input values goes to zero.
A better measure of the error is the cross-validation
error, which is found by successively leaving out well
values and comparing their correct values to the
predicted value.
We will use the standard deviation of the cross-
validation error as our measure of success.
The next two slides show the different errors for the
kriging example.
14-17
Kriging error variance
The error variance map. As expected the error is small at the wells
and gets larger away from the wells.
14-18
Kriging cross-validation error
The cross-validation error map, displaying the absolute error at each
well in % porosity. The standard deviation is 3.25%.
14-19
Cokriging
Cokriging is an extension of kriging, in which the
original equation is extended to include a weighted
sum of secondary (seismic) values.
In collocated cokriging, only the seismic value at the
output location is used (we need 3D seismic data).
As with kriging, an error variance plot can be made
for the cokriged map.
However, we will again use cross-validation as our
measure of accuracy.
The result of applying cokriging to our dataset is
shown in the next slide.
14-20
Cokriged result
This map shows the collocated cokriging result, using impedance as the
secondary variable. Note the imprint of the kriged map. The standard
deviation of the validation error is 2.91%, better than for kriging.
Channel
14-21
The multiattribute transform
The multiattribute transform involves combining
multiple attributes to predict a reservoir parameter
such as porosity, by training at the well locations.
This method can be applied to seismic volumes,
using each well log sample in a window. In this
study, we will use averaged map slices and well
log values.
We will use multilinear regression as our analysis
method.
In the first stage we train the multiattribute
transform.
In the second stage we apply the results of the
training to the multiple map attributes.
14-22
1 1
A w
value log L
m m
A w
The multiattribute map transform
Y
X
2 2
A w
Attribute map 1
Attribute map 2
Attribute map M
This figure shows the multiattribute map transform approach in
schematic form. We need to compute the weights w
i
which, when
multiplied by the attribute values, will produce the log value.
14-23
Multilinear regression
In the multiattribute transform using multilinear
regression, we compute M+1 weights such that the log
value L(X,Y) at a particular map value is a weighted
sum of M attributes A
i
:
The solution to this problem can be found by using a
standard least-squares technique.
A key problem is deciding which attributes to use.
Another important consideration is which of the
attributes are statistically significant
) y , x ( A w ) y , x ( A w w ) y , x ( L
M M 1 1 0

14-24
Attribute slices
We have already seen one of the slices that will be used
in our multiattribute transform: the impedance slice.
The next three slides will show the other slices used.
Each attribute (except trace length) was derived from
the seismic volume by taking an RMS average over a
10 msec window below the picked top of sand.
The following attribute slices were extracted:
Seismic amplitude
Amplitude envelope
Instantaneous phase
Instantaneous frequency
Integrated seismic trace
Trace length the total length of the trace over the window
14-25
Seismic amplitude slices
The map in (a) shows the RMS average of the seismic amplitude over
a 10 ms window below the channel top event, whereas the map in (b)
shows the RMS average of the amplitude envelope over the same
window. Notice that the two slices are very close in appearance.
(a) Seismic amplitude slice. (b) Amplitude envelope slice.
14-26
Instantaneous phase and frequency slices
The map in (a) shows the RMS average of the instantaneous phase over
a 10 ms window below the channel top event, whereas the map in (b)
shows the RMS average of the instantaneous frequency over the same
window.
(a) Instantaneous phase slice. (b) Instantaneous frequency slice.
14-27
Integrated trace and trace length slices
The map in (a) shows the RMS average of the integrated trace over a 10
ms window below the channel top event, whereas the map in (b) shows
the total trace length over the same window.
(a) Integrated trace slice. (b) Total trace length slice.
14-28
Multilinear regression
We are now ready to compute the multilinear
regression map, which will be a linear combination of
the previous maps.
We will also include a non-linear option by computing
transforms (inverse, square root, etc) of the data.
First, we will look at the correlation coefficients between
the wells and each of the attribute slices.
We then compute the best combination of attributes
using a technique called step-wise regression.
Finally, we decide which attributes are significant using
a validation technique in which the target well is left out
in jackknife fashion.
14-29
Correlation coefficients for all the slices
This table shows the correlation coefficients between the well
porosity values and all of the attribute slices, sorted by decreasing
correlation coefficient.
14-30
Validation error plot
This is the validation error for the 5 attributes used in the multiattribute
process. The red line leaves out the target well and shows that only the
first 3 attributes should be used.
14-31
Validation error and weights
The table at the top shows the numerical values from the previous slide
and also shows that the best non-linear fit is between the square root of
porosity and the inverse attributes. The bottom table shows the weights.
Only use
the first
three
attributes
based on
validation
error.
Weights
14-32
Multilinear regression result
This map is the result of applying the multilinear regression
weights, shown in the previous slide, to the attributes. Note that
the result is in pseudo-porosity.
Channel
14-33
New crossplot
This is the new crossplot between the well porosity and the
pseudo-porosity from the multiattribute transform. Note that the
correlation coefficient has gone up to 0.81.
14-34
Combining multilinear regression with
geostatistics
We are now at the final stage of the process, and will
combine the multilinear regression result with the well
values using geostatistics.
That is, the multiattribute transform will replace the
inversion slice as the secondary variable.
The first step is to re-compute the seismic to seismic
variogram.
We will then compute cokriging result.
Finally, we will do a statistical analysis of the results.
14-35
New seismic variogram
This is the seismic variogram used in the final cokriging process.
Note that this is an exponential fit, rather than the spherical fit used in
the earlier variograms.
14-36
Cokriging with the multiattribute transform
This map shows the result of applying the collocated cokriging process
using the multiattribute transform as the secondary variable. The
standard deviation of the validation error is now 2.3%.
Channel
14-37
Map review
Note the increase in geological
information as we move from
(a) the kriged map with wells
alone to (b) cokriging with
inversion, and finally to (c)
cokriging with the multiattribute
transform. However, all three
maps match the wells.
(b) Cokriging with impedance, Std. Dev.=2.91 %
(c) Cokriging with multiple attributes,
Std. Dev.= 2.33 %
(a) Kriging, Std. Dev.= 3.25 %
14-38
Conclusions
In this section of the course, we have combined
geostatistics with multiattribute map analysis.
Traditional cokriging uses a single secondary attribute.
In our case, using impedance as this attribute showed
a strong imprint from the wells.
Using a multiattribute transform, we were able to get a
better fit between the wells and the porosity map.
After a second pass of cokriging, the final map was
more realistic from an exploration point of view.
Statistically, the standard deviation of the cross-
validation error was smallest when we used cokriging
with the multiattribute transform.

Das könnte Ihnen auch gefallen